pyobjconfig 0.1.4

An argparse+pydantic-based configuration system for Python.

pyobjconfig

pyobjconfig is a module suite designed with Machine Learning (ML) experiments in mind, but should work for a broad range of hierarchical configurations.

Example usage

Pyobjconfig supports configuring nested objects, such that each object has control over its own configuration:

import argparse
import pyobjconfig as poc

# Any class which plugs into pyobjconfig should derive from `ConfigurableObject`
class Child(poc.ConfigurableObject):
    # To register configuration options for this class, add a `config` member
    # which derives from a pydantic model.
    class config(poc.PydanticBaseModel):
        # Pydantic handles basic type validation and allows default values to be
        # set
        inner: str = 'hello'

class Base(poc.ConfigurableObject):
    class config(poc.PydanticBaseModel):
        hello: str = 'hi'

    # Assigning a class deriving from `ConfigurableObject` as a member of a
    # class definition results in a nested object, which will be instantiated
    # using parameters from the command line.
    child = Child

ap = argparse.ArgumentParser()
Base.argparse_setup(ap)

args = ap.parse_args(['--child-inner', 'beep']).__dict__
obj = Base.argparse_create(args)
print(obj.config.hello)  # Prints 'hi'
print(obj.child.config.inner)  # Prints 'beep'

ap.print_help()
# usage: ipython [-h] [--hello HELLO] [--child-inner CHILD-INNER]
#
# optional arguments:
#   -h, --help            show this help message and exit
#   --hello HELLO         Default: hi
#   --child-inner CHILD-INNER
#                         Default: hello

Switches are also supported, such that the class (or value) of a child member may be changed at initialization time. Additionally, the --help for the argparse parser will change to reflect the choice of child:

import pyobjconfig as poc
class A(poc.ConfigurableObject):
    pass
class B(poc.ConfigurableObject):
    pass
class Base(poc.ConfigurableObject):
    child = poc.ConfigurableSwitch({
        'a': A,
        'b': B,
        'none': None,
    }, default='none')

Experimentally, environment variables are supported. Not yet well tested, but calling argparse_create(args, env='PREFIX') will allow e.g. PREFIX_CHILD=a to be specified in the environment. Arguments on the command line take precedence.

import argparse
import pyobjconfig as poc

Usage with pytorch

Pytorch overrides __setattr__ in a way that is incompatible with the ConfigurableObject class provided by pyobjconfig. To work around this, use the pyobjconfig.torch.ConfigurableModule as a drop-in replacement for torch.nn.Module.

Changelog

• 2023-04-21 - v0.1.4. argparse.ArgumentParser has type specified for each argument now, so local equality comparisons against previous argparse_hparams() before object instantiation will work.
• 2023-04-18 - v0.1.3. Disallow None as parameter values because it serializes poorly, and disallow abbreviated parameter matches to avoid confusion.
• 2021-01-14 - v0.1.2. Support lists on the command line.